1
PithTrain: A Compact and Agent-Native MoE Training System
提出紧凑且原生支持Agent的MoE训练系统,降低架构演进成本,适配前沿大模型需求。
arXiv:2605.31463v1 Announce Type: cross Abstract: Mixture-of-Experts (MoE) has become the dominant architecture for frontier language models. To meet …